Sound–meaning association biases evidenced across
thousands of languages
Damián E. Blasia,b,c,1, Søren Wichmannd,e, Harald Hammarströmb, Peter F. Stadlerc,f,g, and Morten H. Christiansenh,i
a
Department of Comparative Linguistics and Psycholinguistics Laboratory, University of Zürich, CH-8006 Zurich, Switzerland; bDepartment of Linguistic and
Cultural Evolution, Max Planck Institute for the Science of Human History, 07745 Jena, Germany; cDiscrete Biomathematics Group, Max Planck Institute for
Mathematics in the Sciences, 04103 Leipzig, Germany; dUniversity of Leiden, 2311 BV Leiden, The Netherlands; eKazan Federal University, Kazan, Russia,
420000; fInterdisciplinary Center for Bioinformatics, Department of Computer Science, University of Leipzig, 04107 Leipzig, Germany; gSanta Fe Institute, Santa Fe,
NM 87501; hDepartment of Psychology, Cornell University, Ithaca, NY 14853; and iInteracting Minds Centre, Aarhus University, 8000 Aarhus C, Denmark
Edited by Anne Cutler, University of Western Sydney, Penrith South, NSW, Australia, and approved July 25, 2016 (received for review April 13, 2016)
It is widely assumed that one of the fundamental properties of
spoken language is the arbitrary relation between sound and
meaning. Some exceptions in the form of nonarbitrary associations have been documented in linguistics, cognitive science, and
anthropology, but these studies only involved small subsets of the
6,000+ languages spoken in the world today. By analyzing word
lists covering nearly two-thirds of the world’s languages, we demonstrate that a considerable proportion of 100 basic vocabulary
items carry strong associations with specific kinds of human
speech sounds, occurring persistently across continents and linguistic lineages (linguistic families or isolates). Prominently among
these relations, we find property words (“small” and i, “full” and p
or b) and body part terms (“tongue” and l, “nose” and n). The areal
and historical distribution of these associations suggests that they
often emerge independently rather than being inherited or borrowed. Our results therefore have important implications for the
language sciences, given that nonarbitrary associations have been
proposed to play a critical role in the emergence of cross-modal
mappings, the acquisition of language, and the evolution of our
species’ unique communication system.
|
linguistics cognitive sciences
sound symbolism
| language evolution | iconicity |
A
lthough there is substantial debate in the language sciences
over how to best characterize the features of spoken language, there is nonetheless a general consensus that the relationship between sound and meaning is largely arbitrary (1–3).
Plenty of exceptions exist, however, within individual languages.
For instance, ideophones—a class of words found in many languages—convey a communicative function (or meaning) through
the depiction of sensory imagery (4). In the Mel language Kisi
Kisi (spoken in Sierra Leone), hábá means “(human) wobbly,
clumsy movement,” and hábá-hábá-hábá “(human) prolonged,
extreme wobbling”; here, repetition serves as a way to convey the
meaning of intensity. More generally, the resemblance between
certain aspects of the acoustic basis of speech and their referents,
“iconicity,” is the most researched and well-known case of nonarbitrary associations between sound and meaning (5, 6). “Systemacity,” in contrast, refers to (statistical) regularities that are
common to particular set of words, created by historical contingencies and analogical processes (5). For example, word-initial gl- in English evokes the idea of a visual phenomenon (as in
glare, glance, glimmer) (7). At a larger scale, there is evidence that
the phonological properties of whole morphosyntactic classes of
words (like verbs and nouns) are distinct in several languages (8).
The evidence of recurring regularities in sound–meaning
mappings across multiple languages is considerably more modest, despite its potential importance for fundamental questions
about language evolution and the role of basic perceptual biases
in cognition. For example, certain shape–sound associations—
known as the bouba-kiki effect (9–11)—are believed to rely on
the ability that humans [and perhaps also other primate species
(12)] have for associating stimuli across different modalities (13).
10818–10823 | PNAS | September 27, 2016 | vol. 113 | no. 39
Other plausible sources of cross-linguistic associations include,
for instance, the relationship across many animal species between vocalization frequency and animal size (14), the mimicry
of referents via unconscious mouth gesturing (15), and the persistence of vestiges of a conjectured early human language (16).
Experimental studies support the hypothesis that humans are
indeed sensitive to such associations. It has been demonstrated
several times that participants perform above chance when asked
to pair up words with opposite meanings (antonyms) in languages unknown to them (17), and that English speakers might
even be able to decide on the concreteness of words from languages to which they have not been exposed (18). However, this
evidence for nonarbitrary sound–meaning associations pertains
only to narrow pockets of the vocabulary, making it unclear
whether a more general pressure toward arbitrariness may
overpower such potential biases when considering a more semantically diverse selection of the vocabulary (2, 19).
A further issue with current studies of nonarbitrariness in sound–
meaning correspondences is that, save for a single exception (20),
cross-linguistic corpus studies of nonarbitrary associations have
tended to rely on a small number of languages (maximally 200) and
focusing on small semantically restricted sets of words, ranging from
phonation-related organs (21) to South American animals (15), to
spatial orientation (demonstratives) (14, 22), repair initiators
(like huh? in English) (23), and the conceptualization of magnitude in Australian languages (24). These studies involve confirmatory analyses, aiming to test specific hypotheses regarding
sound–meaning correspondences; as a consequence, they are
guided by a priori intuitions or indirectly by findings from other
disciplines. These limitations may help explain, at least in part,
why language scientists typically consider nonarbitrary associations
to be marginal phenomena that may only apply to small, strictly
Significance
The independence between sound and meaning is believed to be a
crucial property of language: across languages, sequences of different sounds are used to express similar concepts (e.g., Russian
“ptitsa,” Swahili “ndege,” and Japanese “tori” all mean “bird”).
However, a careful statistical examination of words from nearly
two-thirds of the world’s languages reveals that unrelated languages very often use (or avoid) the same sounds for specific
referents. For instance, words for tongue tend to have l or u,
“round” often appears with r, and “small” with i. These striking
similarities call for a reexamination of the fundamental assumption
of the arbitrariness of the sign.
Author contributions: D.E.B., S.W., H.H., and M.H.C. designed research; D.E.B. performed
research; D.E.B. analyzed data; and D.E.B., S.W., H.H., P.F.S., and M.H.C. wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission.
1
To whom correspondence should be addressed. Email: [email protected].
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1605782113/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1605782113
Fig. 1. Geographic distribution of the 6,452 word lists from the ASJP database (25). Colors distinguish different linguistic macroareas, regions with
relatively little or no contact between them (but with much internal contact
between their populations). These are North America (orange), South
America (dark green), Eurasia (blue), Africa (green), Papua New Guinea and
the Pacific Islands (red), and Australia (fuchsia).
Blasi et al.
PNAS | September 27, 2016 | vol. 113 | no. 39 | 10819
ANTHROPOLOGY
Testing Associations on a Global Scale
The availability of a large collection of word lists allows us to
search for statistically robust associations in an unsupervised,
theory-neutral manner. The data consist of 28–40 lexical items
from 6,452 word lists, with a subset of 328 word lists having up to
100 items (25). Words are transcribed into a phonologically
simplified system consisting of 34 consonant and 7 vowels, which
we refer to collectively as “symbols” (Table S1). These words
belong to what is often referred to as “basic vocabulary,” including
for instance pronouns, body part terms, property words, motion
verbs, and nouns describing natural phenomena (26). The word
lists include both languages and dialects, spanning 62% of the
world’s languages and about 85% of its lineages (Fig. 1). A lineage
is a maximal set of languages that can be shown to have a common
ancestor. Such a set may have only one member (an isolate) or
multiple members (a family).
Regarding the classification of languages, the Glottolog genealogical classification is preferable over other available alternatives because it is the only one to classify every living or extinct
language while providing brief pointers to justifications for all
choices taken—however, a less conservative independent classification was used additionally in the main test (see below). We
stratify languages geographically by dividing the world’s landmass into six largely independent linguistic macroareas: North
America, South America, Eurasia, Africa, Greater New Guinea,
and Australia—these regions have a history of attested contact
within them but little contact between them in prehistorical
times (27). To guarantee that only truly global associations were
selected, we screened the sound–meaning associations, keeping
only those where the concept and symbol were attested in languages from at least 10 different lineages and found in no less
than three different macroareas.
We aim to capture robust and widespread tendencies in sound–
meaning associations, where “tendency” should be understood as a
systematic bias in the frequency with which certain words tend to
carry specific symbols in contrast to their baseline occurrence in
other words. Crucially, a strong tendency does not imply that a
signal has an extremely high frequency of occurrence, and conversely a very frequent sound–meaning co-occurrence is not
sufficient evidence to discount chance. Importantly, whatever
advantage a sound–meaning pairing might confer in terms of
learning or processing, it has to be considered in the context of a
myriad of competing factors that shape the phonetic and phonological fabric of words, from articulatory production costs (28)
to systemic constraints due to the similarity with other lexical
elements (29).
Our statistical approach consists in a series of tests where the
presence of a symbol in a word is contrasted against a suitable
subset of other words, and then the bias is evaluated across
lineages. To begin, we calculate, for each concept and symbol, a
genealogically balanced average ratio of the times they co-occur
in a word of a language for which both symbol and concept are
attested. We simulated the same quantity based on the rest of the
concepts and compared it with the previously computed quantity
(Materials and Methods). The associated P value roughly estimates the chance of finding the same or more extreme (genealogically balanced) average by picking any word other than the
target one. Notice that this includes both recurring sound–
meaning pairings as well as its complement, sound–meaning
associations that are observed less often than expected given our
null model.
Crucially, a sequence of tests need to be applied to ensure that
potential associations are not statistical artifacts (see Materials
and Methods and SI Materials and Methods for more details).
First, we used two independent worldwide language classifications with contrasting degrees of conservativeness (30, 31).
Second, we controlled the false discovery rate at a 5% expected
level of false positives (for both classifications independently) so
as to avoid an inflated number of associations due to multiple
comparisons.
Third, word length is trivially correlated with the chance of
finding any particular symbol. There is considerable variance in
the (genealogically balanced) length of the words in our dataset,
with some pronouns, negation, and basic verbs (like say and give)
consisting only of about three symbols on average, whereas the
length of some color words and body part terms contain is over
five (Fig. S1). We filter out associations that also emerge when
all of the symbols of all of the words of each language are randomly permuted while keeping word lengths fixed.
Fourth, besides the mere number of symbols, word length
might be a confound due to the fact that different phonotactic
restrictions might apply accordingly. For instance, in a language
that only allows consonant–vowel structures and also prohibits
the presence of word-initial liquids, no monosyllabic words will
carry liquids. To remedy this, we performed a test similar to the
first one described but this time comparing words only with the
length-matched equivalents of different concepts.
Finally, to filter out associations due to areal contact or unresolved genealogy, we looked for association that could be detected within the linguistic macroareas independently. Thus, we
restricted our attention to associations that passed all these
statistical controls and for which a bias consistent with the
worldwide trend could be found in at least three macroareas,
with no single area showing a bias in the opposite direction.
It should be noted that the overall testing scheme is conservative and that it is likely to have a large false-negative rate. Also
working against our analyses is the fact that the core set of
concepts we use was originally gathered due to their exceptional
phylogenetic persistence and resistance to borrowing, thus rendering them less likely to be adapted to potential functional
biases that might underlie specific sound–meaning associations.
Moreover, it is not clear a priori whether the granularity of our
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
circumscribed regions of the vocabulary (3). In this paper, we
therefore conduct a comprehensive set of analyses involving a semantically diverse set of words from close to two-thirds of the
world’s languages.
phonetic descriptions is sufficiently fine to capture widespread
sound–meaning relations—for instance, the opposition between
voiced and unvoiced consonants and between rounded and
unrounded in vowels have been suggested to bear importance for
sound–symbolism (22, 32), but each feature pair are usually
conflated under a single symbol in the database. For these reasons, the associations found in our analyses should be regarded
as providing a lower-bound estimate of the presence of nonarbitrariness in sound–meaning pairings.
Strong Worldwide Associations
Our analysis detected 74 (positive and negative) sound–meaning
associations, involving 30 concepts and 23 symbols. All of these
associations are referred to as “signals” (Table 1; more detail is
provided in Tables S2 and S3).
Signals will be described in terms of the most relevant information about them: the frequency of the symbol in the words
corresponding to the concept (p), the ratio between that frequency
and the frequency in other words (RR), the number of lineages
that were analyzed for the global association (nl), and the ratio
between the number of areas where the association was independently found and the total number of tested areas (as =at).
Some concepts are associated with more than one signal.
These are expected to be correlated; across languages, it is often
observed that there are preferences or restrictions with regard to
the co-occurrence of symbols within one and the same word for
either diachronic or synchronic phonotactic reasons. As an example, it is known that high front vowels trigger palatalization
(33), so it is therefore not surprising that the voiceless palatoalveolar affricate C appears with i in the signals of small. In a set
Table 1. Summary of signals found in the ASJP database
Concept
Positive symbol
Negative symbol
Ash
Bite
Bone
Breasts
Dog
Drink
Ear
Eye
Fish
Full
Hear
Horn
I
Knee
Leaf
Name
Nose
One
Red
Round
Sand
Skin
Small
Star
Stone
Tongue
Tooth
Water
We
You
u
k
k
um
s
—
k
—
a
pb
N
kr
5
oupkq
bpl
i
un
tn
r
r
s
—
iC
z
t
eEl
—
—
n
—
—
—
y
ahr
t
a
—
a
—
—
—
—
upbtsrl
—
—
op
a
—
—
—
—
mn
—
—
—
uk
bm
t
pls
uoptdqsrl
Positive and negative signals are those that have frequency significantly
larger and smaller than expected.
10820 | www.pnas.org/cgi/doi/10.1073/pnas.1605782113
of testable pairs of signals (Materials and Methods), signals
sharing a concept tend to be significantly associated in about
41% of the time, against only 8% of signals involving different
concepts (Table S4).
The signals found in our analysis show a mixture of wellknown and new associations. In line with the considerable literature on magnitude sound symbolism, the concept small was
found to be associated with the high front vowel i (RR = 1.58,
P = 0.61, nl = 78, as =at = 3/5), consistent with findings linking
vowel height quality and size (14, 17), and with the palatal
consonant C (RR = 5.12, P = 0.41, nl = 61, as =at = 3/4), also in
agreement with previous work (14, 24).
We also observed a strong association between round and r
sounds (RR = 2.48, P = 0.37, nl = 56, as =at = 4/5). Although most
recent research has emphasized the role of consonants in shape–
sound meaning associations like this (34, 35), the usual hypothesis
in this direction concerned the correlation between vowel roundedness and round objects (11)—association that appears as a
tendency in our analyses without reaching the minimum statistical
threshold established before. Both small and round have been
linked to the phenomenon of cross-modal mapping (10, 13, 36).
Another property word, full, is endowed with a pair of signals involving voiced (RR = 1.91, P = 0.22, nl = 213, as =at = 4/6) and
unvoiced bilabial stops (RR = 2.11, P = 0.13, nl = 231, as =at = 5/6).
Some of the strongest signals found correspond to body parts.
Tongue was very strongly associated with the lateral “l” (RR =
2.77, P = 0.41, nl = 280, as =at = 6/6) and the mid and low front
vowels e (RR = 1.54, P = 0.11, nl = 322, as =at = 5/6) and E (RR =
1.73, P = 0.11, nl = 164, as =at = 4/6). Nose was found to be associated most strongly with the alveolar nasal n (RR = 1.47, P =
0.35, nl = 334, as =at = 4/6) and the high back vowel u (RR = 1.38,
P = 0.35, nl = 325, as =at = 4/6). The link between nose and nasality
has been noted previously (37), in particular in reference to the
conjecture that body part terms used in phonation makes use of
the distinctive qualities provided by the relevant organ (21).
Breasts was associated with the bilabial nasal consonant m
(RR = 1.63, P = 0.32, nl = 320, as =at = 4/6) and the high back
vowel u (RR = 1.46, P = 0.37, nl = 317, as =at = 4/6). Similar
associations were found in the nursery terms for mother, a concept with which it often colexifies. It has been suggested that this
might be due to the mouth configuration of suckling babies or to
the sounds feeding babies produce (38, 39).
Although this study lends support to a number of associations
that were either elicited in experiments or conjectured based on
a much smaller number of languages, it also provides telling
negative evidence on others. Together with the association between high front vowels and the concept of small, there has been
reports on a connection between back low vowels and the notion
of big (22). However, big (nl = 73) and large (nl = 74) and o did
not show any relevant signature of association in our sample at
the global level. Similarly, an analogous front/back vowel opposition has been proposed to hold between proximal and distal
pronouns—the purported explanation being that proximal referents tend to be small, whereas distal referents are usually large
(22). The concepts this (nl = 71) and that (nl = 74), however, do
not show any associations with i and o (respectively).
Origins and Nature of the Associations
As discussed in the previous sections, there are multiple theories
that attempt to elucidate why humans find that some sounds are
more convenient or salient in association with certain meanings.
How these hypothesized mechanisms lead to the widespread biases
in vocabularies we find here is a complex question that is unlikely
to be fully answered by the inspection of wordlists. Nonetheless, we
can attempt to evaluate some of the potential consequences of
those theories given the coarse level of detail of our data.
Functional advantages might increase the likelihood of signals
being borrowed across languages in contact with one another,
Blasi et al.
Fig. 2. Competing configurations of the spatial distribution of the tested
languages. Blue and fuchsia dots represent languages with and without a
specific signal, respectively. In the panel to the Left, the likelihood of a
language having the signal is correlated with its geographical distance to its
nearest neighbor, and on the Right, there is no spatial structure.
Blasi et al.
Fig. 3. Genealogical trees of languages where leaves are words for specific
referents. In the figure to the Left, cognate classes (depicted as different
shapes) are associated with signal presence (blue shapes), whereas to the
Right there is no such correspondence.
Conclusion
We have demonstrated that a substantial proportion of words in
the basic vocabulary are biased to carry or to avoid specific sound
segments, both across continents and linguistic lineages. Given
that our analyses suggest that phylogenetic persistence or areal
dispersal are unlikely to explain the widespread presence of
these signals, we are left with the alternative that the signals are
due to factors common to our species, such as sound symbolism,
iconicity, communicative pressures, or synesthesia. We expect future research to further elucidate the role and interaction of these
factors in driving the observed sound–meaning association biases,
and to extend the scope of our findings to a broader portion of
the vocabulary.
The outcome of our analyses have consequences for historicalcomparative linguistics, where it has been suggested that there is
a small set of ultraconserved words that are particularly useful
for establishing ancient genealogical relations beyond the limits
of the comparative method (44). However, some of these words
are involved in the signals discovered here: we is associated with
the alveolar nasal, hear with the velar nasal, and ash with the
vowel u. Thus, proposals of far-reaching etymologies based on
words of similar form and meaning should be accompanied by an
evaluation of whether the observed lexical similarities might have
resulted from the kinds of signal discussed in this paper rather
than common inheritance. More generally, even though it is
unclear whether the locus of the emergence of signals is in the
PNAS | September 27, 2016 | vol. 113 | no. 39 | 10821
ANTHROPOLOGY
words than other shared symbols, using a β mixed-regression
model that distinguishes the effects of symbols, concept, and
lineage (SI Materials and Methods). The model is heavily dominated by the effect of lineage, and signal presence (although
significant) has a negligible effect in the opposite direction than
predicted: the genealogically balanced average effect is less than
a 0.5% decrease in similarity for those words sharing a signalrelated symbol compared with those sharing some other symbol.
Consistency in word position is important for establishing
cognacy (45, 46). Further support for the idea that signals are not
residuals of deep history comes from the analysis of the position
within the word in which they occur, in particular whether they
have a clear word-initial bias. All in all, we find that signals do
not have a consistent cross-linguistic preference or dispreference
in this respect beyond well-established cross-linguistic phonotactic patterns, such as the avoidance of liquids or the prevalence
of dorsal and labial stops in word-initial position (47, 48) (SI
Materials and Methods and Table S6).
These results suggest that, although it is possible that the
presence of signals in some families are symptomatic of a particularly pervasive cognate set, this is not the usual case. Hence,
the explanation for the observed prevalence of sound–meaning
associations across the world has to be found elsewhere (49).
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
thus producing spatial diffusion patterns (39) (Fig. 2). The existence of opposing factors obscure definitive inferences in this
direction, however: basic vocabulary items are particularly resistant to borrowing, but unresolved genealogy involving nearby
languages would be confounded with borrowing. In the same
direction, large populations have been claimed to be more efficient at gaining and retaining nonarbitrary sound–meaning associations given a potential functional value (39), which is
coherent with recent evidence from some Austronesian languages showing that larger populations gain new words at a
faster rate (40).
We determined whether present-day log population size and
log distance to the nearest genealogically unrelated language
bearing the (positive) signal are effective predictors for signal
presence, via a mixed-effects logistic model (Table S5 and SI
Materials and Methods). At α = 0.05, log population turned out to
be significant in about one-third of the cases, but the effect was
small and as many times positive as it was negative, which rules
out a consistent role for population. Only one-fifth of the signals
showed sensitivity to the distance of nearest neighbors with signal, with all of the cases having an effect in the predicted direction by our model. On average, and in contrast to the case in
which a language and its signal-bearing nearest genealogically
unrelated neighbor are spoken in exactly the same place, the
probability of finding the signal also in the language drops
by 28%.
From a historical perspective, it has been suggested that
sound–meaning associations might be evolutionarily preserved
features of spoken language (41), potentially hindering regular
sound change (17). Furthermore, it has been claimed that
widespread sound–meaning associations might be vestiges of one
or more large-scale prehistoric protolanguages (16). Tellingly,
some of the signals found here feature prominently in reconstructed “global etymologies” (42, 43) that have been used for
deep phylogeny inference (44). If signals are inherited from an
ancestral language spoken in remote prehistory, we might expect
them to be distributed similarly to inherited, cognate words; that
is, their distribution should to a large extent be congruent with
the nodes defining their linguistic phylogeny (see Fig. 3 for
illustration).
A direct evaluation of this hypothesis is infeasible due to the
absence of etymological dictionaries for all but a few families.
However, it can be tested indirectly given that cognate words are
expected to be more similar to one another than noncognates
(45). We investigated whether the presence of the signal-bearing
symbol was a better indicator of overall form similarity between
invention or historical development of lexical roots, our findings have implications for the study of the dynamics of lexical
phonology.
In summary, our results provide insights into the constraints
that affect how we communicate, suggesting that despite the
immense flexibility of the world’s languages, some sound–meaning
associations are preferred by culturally, historically, and geographically diverse human groups.
Materials and Methods
Basic Vocabulary Word Lists. The dataset used for this study is drawn from
version 16 of the Automated Similarity Judgment Program (ASJP) database
(25). ASJP comprises 6,895 word lists from around 62% of the world’s languages, covering 85% of families, isolates, and unclassified languages [using
the Ethnologue (50) for these statistics]. After removing artificial languages,
pidgins, and creoles, and varieties whose ISO-639-3 code cannot be confirmed, the number goes down to 6,447 word lists, corresponding to 4,298
different languages and 359 lineages. The database was not constructed for
the specific purpose of studying sound symbolism, but rather for identifying
genealogical relations among languages. For this reason, it generally consists of the 40-item subset of the 100-item so-called Swadesh list (51) that are
assumed to remain stable as languages diverge into different lineages over
time (52). Of these word lists, 328 additionally contain the remaining 60
Swadesh lists items.
Words are rendered in a unified transcription system, which facilitates
cross-linguistic comparison but also ignores phonetic details such as vowel
length, nasalization, tones, and retroflexation. Vowel quality distinctions are
merged into seven categories (high front, mid front, low front, high-mid
central, low central, high back, and midlow back) (see ref. 53 for a discussion
of the system).
Each 40-item word list provides translational equivalents, when available,
for the following items: blood, bone, breast, come, die, dog, drink, ear, eye,
fire, fish, full, hand, hear, horn, I, knee, leaf, liver, louse, mountain, name,
new, night, nose, one, path, person, see, skin, star, stone, sun, tongue, tooth,
tree, two, water, we, and you (sg). The additional Swadesh list items contained in some of the word lists are as follows: all, ash, bark, belly, big, bird,
bite, black, burn, claw, cloud, cold, dry, earth, eat, egg, feather, flesh, fly,
foot, give, good, grease, green, hair, head, heart, hot, kill, know, lie, long,
man, many, moon, mouth, neck, not, rain, red, root, round, sand, say, seed,
sit, sleep, small, smoke, stand, swim, tail, that, this, walk, what, white, who,
woman, and yellow.
Associations Between Symbols and Concepts. The fundamental statistic in our
analysis is pij , the maximum-likelihood estimator (i.e., the sample frequency)
for the probability of finding that concept i has at least one instance of
symbol j, after randomly choosing a lineage, a language within the lineage
and a dialect within the language (if any) in that sequential order. Naturally,
this calculation is restricted to the set of dialects of languages for which the
concept and the phone are attested (which we will refer as Sij ); for each of
those sets, this quantity is formally
pij =
!
jLj
jLk j
jLkl j
1 X 1 X
1 X
π kld
.
ij
jLj k=1 jLk j l=1 jLkl j d=1
The sets L, Lk , and Lkl are the sets of all lineages, languages within lineage k
and dialects of language l within lineage k. π kld
ij is a binary variable that takes
value 1 if there is at least one instance of symbol j in the word for concept i
for dialect d of language l from lineage k (always within the set Sij ) and
0 otherwise.
This computation is conservative in that all languages known to belong to
the same genealogical group influence the aggregated statistics in the same
way regardless of their size, but on the other hand it guarantees the minimum possible bias in the dependence of the languages’ words. To avoid
testing cases whose coverage is insufficiently wide before testing, we evaluated only those associations for which Sij comprises 10 lineages in each of
three different macroareas at least.
Conversely, for each dialect of each language, we calculated the proportion of words other than that associated with i that have symbol j, and we
note this as π kld
−ij , and similarly the genealogical balanced average as p−ij .
These probabilities are used to produce nsim = 1,000 Monte Carlo simulations
of symbol j presence/absence for all of the languages in Sij —the set of p−ij
values resulting from these simulations will be called ζ ij . The purpose is to
compare ζ ij with π ij to answer the question: does symbol j appear much more
(or much less) often when a subset of words referring to concept i is selected
10822 | www.pnas.org/cgi/doi/10.1073/pnas.1605782113
than in a randomly picked set of words from the same languages? The twotailed P value for a particular concept i and symbol j is then as follows (54):
P=
" "
"o
#
!
n"
1
" "
"
"
2 min "x ∈ ζ ij : x ≥ pij ", "x ∈ ζ ij : x ≤ pij " + 1 ,
nsim + 1
where j·j is the cardinality of the set.
The large number of tests performed require a control for type I errors. We
perform a false discovery rate (FDR) analysis fixing the FDR rejection threshold
to 0.05, which means that we will allow no more than 5% of false positives on
average. For this purpose, we use the method described in ref. 55. The basic
idea is that the distribution of P values comes from a mixture of a uniform
distribution (that corresponds to the baseline of tests where no associations
beyond chance are present) and a distribution concentrated near P = 0 of
true positives. The method used here learns the mixture proportion of the
uniform distribution from values P from 1 down to a threshold that is adjusted to reduce the false nondiscovery rate.
This entire procedure was repeated with a different, less conservative,
genealogical classification—the one provided by the World Atlas of Language Structures (WALS) (30). For our analysis, we only considered associations
that were below the defined FDR level according to both classifications. The
fraction of the component of true negatives learned from both classifications
was around 0.65.
Regarding possible confounds due to word length, we performed two
extra tests on those associations that successfully passed the previous test.
First, we repeated the same global test using the Glottolog classification this
time comparing pij with simulations obtained from words of exactly the
same number of symbols in each language (and dialect). Second, for each
language (and dialect) in Sij , n = 1,000 of independent simulations we sampled without replacement as many random symbols from words other than
i up to the length of word i. This effectively produces, for each word i, a
random counterpart equivalent to shuffling all of the symbols corresponding to all of the words of a language while keeping word lengths constant.
Over each of those sets, the same association test based on the Glottolog
classification was performed. In both of these procedures, we imposed a
stricter cutoff: if any of the simulations yield a value of pij equally or more
extreme, we would reject the association as of potential interest.
Finally, for each macroarea with at least 10 independent lineages in Sij , we
analyzed the presence of a significant direction of association as in the main
associations test—computing both empirical and random probabilities using
only the languages of that area—with the difference that we flagged each
macroarea-specific association with P ≤ 0.1. It should be noticed that this
does not imply a softer rejection threshold than in the worldwide case: we
only keep associations that display a bias consistent with the worldwide
trend in at least one-half of the macroareas, with the extra condition that no
macroarea should exhibit a bias in the opposite direction.
To summarize: only associations that successfully satisfied all of the requirements of the overall association test (with Glottolog and WALS classifications independently), the word length and the matched-length tests, and
for which a consistent preference in at least one-half of the macroareas could
be found were considered “signals.”
Association Between Signals. As in the previous case, we analyze sets of
languages for which both the concept and the symbol associated with a pair
of signals was present in at least 10 lineages in each of (at least) three
macroareas. The association between signals—which we will refer to A and B
here—was tested by means of a simple mixed-effects logistic model as
follows:
logitðsignal A presenceÞ = αsignal B presence + αlineage ,
where αsignal A presence is the coefficient related to the presence of signal A,
and αlineage is a random coefficient structured according to lineage. To the
results obtained by comparing all of the pairwise associations between signals belonging to the core 40 words, we applied a threshold on the FDR of
5%. About 12% of the 2,062 cases satisfied this condition. The results of
associations regarding same-concept signals and the genealogically balanced average effect on the presence of signal B on A can be found in
Table S4.
ACKNOWLEDGMENTS. We acknowledge the comments of Bernard Comrie,
Brent O. Berlin, Stephen C. Levinson, Mark Dingemanse, Russell Gray, and
Eric W. Holman. We also thank Jeremy Collins and Stefany Moreno for
assistance with other aspects of the manuscript. H.H.’s research was made
possible thanks to the financial support of the Language and Cognition
Department at the Max Planck Institute for Psycholinguistics, Max-Planck
Blasi et al.
1. Saussure FD (1916) Cours de Linguistique Générale, eds Bally C, Sechehaye A,
Riedlienger A (Payot, Paris).
2. Hockett CF (1960) The origin of speech. Sci Am 203:89–96.
3. Pinker S (1999) Words and Rules: The Ingredients of Language (Perseus Books, New
York).
4. Dingemanse M (2012) Advances in the cross-linguistic study of ideophones. Lang
Linguist Compass 6(10):654–672.
5. Dingemanse M, Blasi DE, Lupyan G, Christiansen MH, Monaghan P (2015) Arbitrariness, iconicity, and systematicity in language. Trends Cogn Sci 19(10):603–615.
6. Monaghan P, Shillcock RC, Christiansen MH, Kirby S (2014) How arbitrary is language?
Philos Trans R Soc Lond B Biol Sci 369(1651):20130299.
7. Bergen BK (2004) The psychological reality of phonaesthemes. Language 80:290–311.
8. Monaghan P, Christiansen MH, Chater N (2007) The phonological-distributional coherence hypothesis: Cross-linguistic evidence in language acquisition. Cognit Psychol
55(4):259–305.
9. Köhler W (1929) Gestalt Psychology (Liveright, New York).
10. Ramachandran VS, Hubbard EM (2001) Synaesthesia—a window into perception,
thought and language. J Conscious Stud 8(12):3–34.
11. Maurer D, Pathman T, Mondloch CJ (2006) The shape of boubas: Sound-shape correspondences in toddlers and adults. Dev Sci 9(3):316–322.
12. Ludwig VU, Adachi I, Matsuzawa T (2011) Visuoauditory mappings between high
luminance and high pitch are shared by chimpanzees (Pan troglodytes) and humans.
Proc Natl Acad Sci USA 108(51):20661–20665.
13. Cuskley C, Kirby S (2013) Synaesthesia, cross-modality and language evolution. Oxford
Handbook of Synaesthesia, eds Simner J, Hubbard EM (Oxford Univ Press, Oxford,
UK), pp 869–907.
14. Hinton L, Nichols J, Ohala JJ, eds (2006) Sound Symbolism (Cambridge Univ Press, New
York).
15. Berlin B (2006) The first congress of ethnozoological nomenclature. J R Anthropol Inst
12(s1):S23–S44.
16. Imai M, Kita S (2014) The sound symbolism bootstrapping hypothesis for language
acquisition and language evolution. Philos Trans R Soc Lond B Biol Sci 369(1651):
20130298.
17. Nuckolls JB (1999) The case for sound symbolism. Annu Rev Anthropol 28:225–252.
18. Reilly J, Hung J, Westbury C (2016) Non arbitrariness in mapping word form to
meaning: Cross-linguistic formal markers of word concreteness. Cogn Sci, 10.1111/
cogs.12361.
19. Monaghan P, Christiansen MH, Fitneva SA (2011) The arbitrariness of the sign:
Learning advantages from the structure of the vocabulary. J Exp Psychol Gen 140(3):
325–347.
20. Wichmann S, Holman EW, Brown CH (2010) Sound symbolism in basic vocabulary.
Entropy 12:844–858.
21. Urban M (2011) Conventional sound symbolism in terms for organs of speech: A crosslinguistic study. Folia Linguist 45(1):199–214.
22. Johansson N, Zlatev J (2013) Motivations for sound symbolism in spatial deixis: A
typological study of 101 languages. Public J Semiotics 5(1):3–20.
23. Dingemanse M, Torreira F, Enfield NJ (2013) Is “huh?” a universal word? Conversational infrastructure and the convergent evolution of linguistic items. PLoS One 8(11):
e78273.
24. Haynie H, Bowern C, Lapalombara H (2014) Sound symbolism in the languages of
Australia. PLoS One 9(4):e92852.
25. Wichmann S, et al. (2013) The ASJP Database (Version 16). Available at asjp.clld.org/.
Accessed July 2, 2013.
26. Tadmor U, Haspelmath M, Taylor B (2010) Borrowability and the notion of basic
vocabulary. Diachronica 27(2):226–246.
27. Hammarström H, Donohue M (2014) Some principles on the use of macro-areas in
typological comparison. Lang Dynam Change 4(1):167–187.
28. Napoli DJ, Sanders N, Wright R (2014) On the linguistic effects of articulatory ease,
with a focus on sign languages. Language 90(2):424–456.
29. Vitevitch MS, Luce PA (2016) Phonological neighborhood effects in spoken word
perception and production. Annu Rev Linguistics 2:75–94.
30. Haspelmath M, Dryer MS, Gil D, Comrie B (2005) The World Atlas of Language
Structures (Oxford Univ Press, Oxford, UK).
31. Nordhoff S, Hammarström H, Forkel R, Haspelmath M (2013) Glottolog 2.1. Available
at glottolog.org. Accessed July 2, 2013.
32. Lockwood G, Dingemanse M (2015) Iconicity in the lab: A review of behavioral, developmental, and neuroimaging research into sound-symbolism. Front Psychol 6:1624.
33. Bateman N (2011) On the typology of palatalization. Lang Linguist Compass
5(8):588–602.
34. Nielsen A, Rendall D (2011) The sound of round: Evaluating the sound-symbolic role
of consonants in the classic Takete-Maluma phenomenon. Can J Exp Psychol 65(2):
115–124.
35. Fort M, Martin A, Peperkamp S (2015) Consonants are more important than vowels in
the Bouba-kiki effect. Lang Speech 58(Pt 2):247–266.
36. Bankieris K, Simner J (2015) What is the link between synaesthesia and sound symbolism? Cognition 136:186–195.
37. Greenberg JH, Ferguson CA, Moravcsik EA, eds (1978) Universals of Human Language:
Phonology (Stanford Univ Press, Stanford, CA), Vol 2.
38. Jakobson R (1960) Why “mama” and “papa.” Perspectives in Psychological Theory,
Dedicated to Heinz Werner, eds Kaplan B, Wapner S (International Universities Press,
New York), pp 124–134.
39. Traunmüller H (1994) Sound symbolism in deictic words. Tongues and Texts Unlimited. Studies in Honour of Tore Jansson on the Occasion of his Sixtieth Anniversary,
eds Janson T, Aili H, af Trampe P (Stockholms Universitet, Institutionen för Klassiska
Språk, Stockholm), pp 213–234.
40. Bromham L, Hua X, Fitzpatrick TG, Greenhill SJ (2015) Rate of language evolution is
affected by population size. Proc Natl Acad Sci USA 112(7):2097–2102.
41. Nygaard LC, Cook AE, Namy LL (2009) Sound to meaning correspondences facilitate
word learning. Cognition 112(1):181–186.
42. Ruhlen M (1994) On the Origin of Languages: Studies in Linguistic Taxonomy (Stanford
Univ Press, Stanford, CA).
43. Starostin SA, Bronnikov Y (2009) Languages of the World Etymological Database, Part
of the Tower of Babel–Evolution of Human Language Project. Available at starling.
rinet.ru/cgi-bin/main.cgi?flags=eygtnnl. Accessed May 25, 2015.
44. Pagel M, Atkinson QD, S Calude A, Meade A (2013) Ultraconserved words point to
deep language ancestry across Eurasia. Proc Natl Acad Sci USA 110(21):8471–8476.
45. Steiner L, Stadler PF, Cysouw M (2011) A pipeline for computational historical linguistics. Lang Dynam Change 1(1):89–127.
46. Jäger G (2015) Support for linguistic macrofamilies from weighted sequence alignment. Proc Natl Acad Sci USA 112(41):12752–12757.
47. Proctor MI (1995) Gestural characterization of a phonological class: The liquids. PhD
dissertation (Yale University, New Haven, CT).
48. MacNeilage PF, Davis BL (2000) On the origin of internal structure of word forms.
Science 288(5465):527–531.
49. Campbell L, Poser WJ (2008) Language Classification (Cambridge Univ Press, New
York).
50. Lewis P, Simons G, Fennig C (2013) Ethnologue: Languages of the World (SIL International, Dallas, TX).
51. Swadesh M (1955) Towards greater accuracy in lexicostatistic dating. Int J Am Linguist
21(2):121–137.
52. Holman EW, et al. (2008) Explorations in automated language classification. Folia
Linguist 42(2):331–354.
53. Brown CH, Holman EW, Wichmann S (2013) Sound correspondences in the world’s
languages. Language 89(1):4–29.
54. North BV, Curtis D, Sham PC (2002) A note on the calculation of empirical P values
from Monte Carlo procedures. Am J Hum Genet 71(2):439–441.
55. Strimmer K (2008) fdrtool: A versatile R package for estimating local and tail areabased false discovery rates. Bioinformatics 24(12):1461–1462.
56. Nichols J, Witzlack-Makarevich A, Bickel B (2013) The AUTOTYP Genealogy and Geography Database: 2013 Release (University of Zürich, Zurich). Available at www.
autotyp.uzh.ch/. Accessed May 29, 2015.
57. Dunn M, Greenhill SJ, Levinson SC, Gray RD (2011) Evolved structure of language
shows lineage-specific trends in word-order universals. Nature 473(7345):79–82.
58. Gray RD, Drummond AJ, Greenhill SJ (2009) Language phylogenies reveal expansion
pulses and pauses in Pacific settlement. Science 323(5913):479–483.
59. Gooskens C, Heeringa W (2004) Perceptive evaluation of Levenshtein dialect distance
easurements using Norwegian dialect data. Lang Var Change 16:189–207.
60. Nerbonne J, Heeringa W, Kleiweg P (1999) Edit distance and dialect proximity. Time
Warps, String Edits and Macromolecules: The Theory and Practice of Sequence
Comparison, eds Sankoff D, Kruskal J (CSLI Press, Stanford, CA), pp v–xv.
61. Smithson M, Verkuilen J (2006) A better lemon squeezer? Maximum-likelihood regression with beta-distributed dependent variables. Psychol Methods 11(1):54–71.
ANTHROPOLOGY
Agreement 295918 and by a subsidy of the Russian Government to support
the Program of Competitive Development of Kazan Federal University. D.E.B.’s
research was funded by the Max Planck Institute International Research
School.
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
Gesellschaft, and European Research Council’s Advanced Grant 269484
(“INTERACT”) to Stephen C. Levinson. S.W.’s research was supported by
the European Research Council under the European Union’s Seventh
Framework Programme (FP7/2007-2013)/European Research Council Grant
Blasi et al.
PNAS | September 27, 2016 | vol. 113 | no. 39 | 10823
Supporting Information
Blasi et al. 10.1073/pnas.1605782113
SI Materials and Methods
Positional Test. We simulate, for each language and signal, random
positions of the relevant signal-associated symbol based on all of
the available positions in the word according to the consonant/
vowel distinction. Concretely, we calculate the number of times
the phone is initial when its simulated counterpart is not, averaging genealogically and respecting the vowel and consonant
template of each word. Then we compare this quantity in the
original word list against n = 1,000 simulations and consider those
cases in which the original bias is larger than 95% of the simulated cases. These results can be observed in Table S6.
Areal and Population Test. For each positive signal we calculated
the great circle distances—i.e., the distance in kilometers of the
shortest geodesic connecting two points in the surface of the
Earth—involving all languages having both the relevant symbol
and concept (but not necessarily the signal) and their nearest
language from a different lineage that has the (positive) signal
(dnn). The hypothesis is that small distance from a language that
has a signal will influence the likelihood of signal presence in a
given language. Only signals belonging to the group of 28–40
better attested concepts were used for the analysis, and only one
dialect per language was chosen. Extinct languages were excluded from the analyses.
For the testing, we used a generalized logistic model with
random effects
!
"
lineage
logitðE½signal presence#Þ = α + βdnn + βdnn
logð1 + dnnÞ
+ βpop logðpopulationÞ + αlineage ,
lineage
where the superscripted coefficients (βdnn and αlineage) are random effects structured according to the lineage. Lineage as a
random intercept is introduced as a means of accounting for
the varying baseline presence of the signals within lineages,
and their presence as random slopes aims to capture the fact
that lineages have spread with different rates across the globe.
The logarithmic transforms aim to reduce the effect of population and distance outliers. P values were estimated through an
asymptotic likelihood ratio test. Apart from the estimated coefficients, we calculated the genealogical balanced mean difference in probability of having a signal for two reference points,
one variable at a time. For population, the difference was calculated between fixing all languages’ populations to 10,000 individuals and a single individual, and for dnn between 1,000 km—
which is roughly the maximum radius of linguistic areas as defined in AUTOTYP (56)—and 0 km (which corresponds to the
situation where both languages as spoken at the same place).
The results can be observed in Table S5.
Word Similarity Test. Ideally, a proper phylogenetic test in the
context of language history would comprise some kind of data
carrying a phylogenetic signal (like cognate sets or collections of
regular sound changes) and a sound evolutionary model that
would lead to a tree or a distribution of trees. Unfortunately, such
trees exist for only a handful of language families (57, 58). Instead,
we approach the question of both phylogenetic stability and
ancestry of signals by analyzing word form similarity, which serves
Blasi et al. www.pnas.org/cgi/content/short/1605782113
as a proxy for cognacy. If it is a correct hypothesis that signals
render words less prone to change and that they are prehistoric
vestiges, then, after controlling for concept, symbol, and lineage,
we would expect to find that the similarity among words is predicted by signals.
The distance between words used here is the Levenshtein distance, which has found several uses in linguistics and often correlates with perceptual, processing, and other meaningful lexical
distances differences (59, 60). The Levenshtein distance between
strings x and y LD(x,y) is defined as the minimum number of edits,
additions, or deletions of characters necessary to make two strings
identical. For instance, “Zultus” and “sulus”—star in Uyghur and
Sakha (two Turkic languages), respectively, have a Levenshtein
distance of 2: a change of “Z” to “s” and the deletion of “t” in the
Sakha word. The normalized Levenshtein distance is simply
l = LDðx, yÞ=maxðjxj, jyjÞ.
For every family with at least six languages and every combination of concept and symbol, we calculated the Levenshtein
distance between all members of two groups: word pairs for a
concept belonging to a combination, and word pairs for a concept
sharing at least one symbol but not the symbol relevant for the
combination. For instance, given a family with three languages
having the forms ana, ena, and ete for the concept “rock,” and
considering the combination rock-n, we will have the two following groups: (ana,ena) and (ena,ete). Families with less than
three distances in any of the groups were excluded from the
analysis.
To summarize the previous information, we calculated, for
each family, the probability of choosing a distance in the signalsharing group and another in the non–signal-sharing group and
finding that the first is smaller than the second [Prðls < l−s Þ]. The
larger this quantity, the more reliable an estimator of word form
similarity the association is.
Then we implemented the following β regression mixed model
with logistic link function and constant precision parameter
X
X
logitðE½Prðls < l−s Þ#Þ =
β i Ii +
β j Ij
concepts
symbols
+ αsignalhood + αlineage ,
where the i and j indexes run over the set of concepts and symbols, respectively; the coefficient “signalhood” indicates whether
the combination of concept and symbol is to be found in Table
S2. “Signalhood” was coded as a single level common to all individual positive signals. αlineage stands for a random intercept
according to lineage. To cope with a few values of Prðls < l−s Þ
identical to 1 (that account for less than 0.5% of the data), we
applied the transformation tðxÞ = ðxðN − 1Þ + 0.5Þ=N to the values
(61). As a way of accounting for the more robust evidence provided by lineages with a large number of distance pairs to be
compared, we included a weight for each observation equal to
the logarithm of the number of such pairs involved—however,
the results did not differ considerably from the unweighted case.
Overall, the model quality is heavily dominated by lineage: 86%
vs. 3% of explained deviance with and without the lineage random effect, respectively.
1 of 7
I
you
water
come
drink
fire
tree
we
name
tooth
louse
eye
see
breasts
hand
bone
die
stone
blood
dog
path
fish
sun
person
nose
skin
hear
leaf
liver
two
one
ear
horn
night
tongue
new
mountain
full
star
knee
20
15
10
5
0
3
4
5
Number of characters
6
3
4
5
Number of characters
6
7
Fig. S1. On the Left, genealogically balanced average of the number of characters for each of the 40 concepts with most coverage in ASJP. The horizontal bars
represent approximate 95% CI for the average. On the Right, distribution of the genealogically balanced average for all of the concepts in ASJP. In both
graphs, the vertical blue bar represents the mean value across all concepts in ASJP.
Blasi et al. www.pnas.org/cgi/content/short/1605782113
2 of 7
Table S1. ASJP symbols and their description
Symbol
p
b
m
f
v
8
4
t
d
s
z
c
n
S
Z
C
j
T
5
k
g
x
N
q
X
Description
Voiceless bilabial stop and fricative
Voiced labial stop and fricative
Bilabial nasal
Voiceless labiodental fricative
Voiced labiodental fricative
Voiceless and voiced dental fricative
Dental nasal
Voiceless alveolar stop
Voiced alveolar stop
Voiceless alveolar fricative
Voiced alveolar fricative
Voiceless and voiced alveolar fricative
Voiceless and voiced alveolar nasal
Voiceless postalveolar fricative
Voiced postalveolar fricative
Voiceless palato-alveolar affricative
Voiced palato-alveolar affricate
Voiceless and voiced palatal stop
Palatal nasal
Voiceless velar stop
Voiced velar stop
Voiceless and voiced velar fricative
Velar nasal
Voiceless and voiced uvular stop
Voiceless and voiced uvular fricative, voiceless and
voiced pharyngeal fricative
Voiceless glottal stop
Voiceless and voiced glottal fricative
Voiced alveolar lateral approximate
All other laterals
Voiced bilabial-velar approximant
Palatal approximant
All varieties of “r sounds”
High front vowel, rounded and unrounded
Mid front vowel, rounded and unrounded
Low front vowel, rounded and unrounded
High and mid central vowel, rounded and unrounded
Low central vowel, unrounded
High back vowel, rounded and unrounded
Mid and low back vowel, rounded and unrounded
7
h
l
L
w
y
r
i
e
E
3
a
u
o
IPA equivalents of the symbols can be found in Tables 1–2 of (53).
Blasi et al. www.pnas.org/cgi/content/short/1605782113
3 of 7
Table S2. Complete list of positive signals found in the ASJP
database
Concept Symbol
Ash
Bite
Bone
Breasts
Breasts
Dog
Ear
Fish
Full
Full
Hear
Horn
Horn
I
Knee
Knee
Knee
Knee
Knee
Leaf
Leaf
Leaf
Name
Nose
Nose
One
One
Red
Round
Sand
Small
Small
Star
Stone
Tongue
Tongue
Tongue
We
u
k
k
u
m
s
k
a
p
b
N
k
r
5
u
o
p
k
q
p
b
l
i
u
n
t
n
r
r
s
i
C
z
t
e
E
l
n
pij
p−ij
σðp−ij Þ
Δ
RR
Lineages
Areal
ratio
0.516
0.438
0.311
0.376
0.326
0.225
0.319
0.613
0.255
0.229
0.199
0.339
0.271
0.129
0.472
0.406
0.218
0.374
0.313
0.232
0.185
0.268
0.474
0.351
0.356
0.266
0.320
0.350
0.371
0.325
0.613
0.416
0.158
0.239
0.339
0.278
0.419
0.380
0.270
0.259
0.223
0.257
0.200
0.128
0.224
0.524
0.121
0.120
0.127
0.222
0.155
0.063
0.256
0.239
0.121
0.226
0.136
0.119
0.124
0.154
0.378
0.255
0.242
0.178
0.248
0.156
0.149
0.126
0.389
0.081
0.063
0.181
0.220
0.161
0.151
0.246
0.043
0.042
0.016
0.018
0.016
0.015
0.017
0.019
0.016
0.016
0.018
0.019
0.019
0.015
0.018
0.017
0.014
0.018
0.027
0.014
0.014
0.016
0.020
0.018
0.016
0.015
0.017
0.037
0.038
0.034
0.043
0.029
0.018
0.015
0.017
0.020
0.017
0.017
0.25
0.18
0.09
0.12
0.13
0.10
0.09
0.09
0.13
0.11
0.07
0.12
0.12
0.07
0.22
0.17
0.10
0.15
0.18
0.11
0.06
0.11
0.10
0.10
0.11
0.09
0.07
0.19
0.22
0.20
0.22
0.33
0.10
0.06
0.12
0.12
0.27
0.13
1.91
1.69
1.39
1.46
1.63
1.76
1.42
1.17
2.11
1.91
1.57
1.53
1.75
2.06
1.84
1.70
1.81
1.66
2.30
1.94
1.48
1.75
1.25
1.38
1.47
1.49
1.29
2.24
2.48
2.58
1.58
5.12
2.52
1.32
1.54
1.73
2.77
1.54
68
73
333
317
320
285
338
327
231
213
182
221
191
136
303
291
278
305
73
290
274
270
320
325
334
343
348
61
56
65
78
61
96
340
322
164
280
325
3/5
3/5
3/6
4/6
4/6
3/5
4/6
3/6
5/6
4/6
3/6
4/6
3/6
4/6
4/6
4/6
5/6
5/6
3/5
3/6
3/6
4/6
3/6
4/6
4/6
3/6
3/6
3/5
4/5
3/5
3/5
3/4
3/5
3/6
5/6
4/6
6/6
3/6
The column “Areal ratio” indicates the ratio between the number of
areas where the signals are independently found with respect the total
number of areas with minimum coverage. RR stands for “risk ratio.” Family
counts come from Glottolog (31).
Blasi et al. www.pnas.org/cgi/content/short/1605782113
4 of 7
Table S3. Complete list of negative signals found in the ASJP
database
Concept Symbol
Bone
Breasts
Breasts
Breasts
Dog
Drink
Eye
I
I
I
I
I
I
I
Name
Name
Nose
Skin
Skin
Tongue
Tongue
Tooth
Tooth
Water
We
We
We
You
You
You
You
You
You
You
You
You
y
a
h
r
t
a
a
u
p
b
t
s
l
r
o
p
a
m
n
u
k
b
m
t
p
l
s
u
o
p
t
d
q
s
r
l
pij
p−ij
σðp−ij Þ
Δ
RR
Lineages
Areal
ratio
0.065
0.422
0.093
0.083
0.106
0.421
0.423
0.116
0.021
0.030
0.079
0.036
0.030
0.061
0.169
0.049
0.391
0.109
0.170
0.164
0.167
0.054
0.130
0.066
0.052
0.064
0.077
0.149
0.165
0.046
0.072
0.045
0.043
0.049
0.053
0.030
0.122
0.524
0.149
0.175
0.182
0.533
0.527
0.262
0.122
0.124
0.181
0.131
0.161
0.177
0.254
0.122
0.524
0.207
0.256
0.264
0.232
0.126
0.205
0.184
0.121
0.160
0.129
0.259
0.246
0.124
0.182
0.129
0.146
0.131
0.180
0.159
0.013
0.020
0.016
0.015
0.015
0.020
0.018
0.018
0.014
0.014
0.016
0.015
0.016
0.015
0.018
0.015
0.019
0.016
0.016
0.017
0.017
0.014
0.016
0.015
0.015
0.016
0.015
0.017
0.017
0.014
0.015
0.015
0.029
0.015
0.016
0.016
−0.06
−0.10
−0.06
−0.09
−0.08
−0.11
−0.10
−0.15
−0.10
−0.09
−0.10
−0.10
−0.13
−0.12
−0.09
−0.07
−0.13
−0.10
−0.09
−0.10
−0.07
−0.07
−0.08
−0.12
−0.07
−0.10
−0.05
−0.11
−0.08
−0.08
−0.11
−0.08
−0.10
−0.08
−0.13
−0.13
0.54
0.81
0.62
0.47
0.58
0.79
0.80
0.44
0.18
0.24
0.44
0.27
0.19
0.35
0.67
0.40
0.75
0.53
0.66
0.62
0.72
0.43
0.63
0.36
0.43
0.40
0.60
0.58
0.67
0.37
0.40
0.35
0.29
0.37
0.29
0.19
312
329
254
290
337
310
357
328
297
276
332
279
277
294
297
283
339
323
329
327
334
282
335
345
288
268
273
316
306
289
322
264
75
271
284
266
3/6
3/6
3/6
3/6
4/6
4/6
4/6
5/6
5/6
4/6
4/6
4/5
6/6
6/6
4/6
3/6
4/6
4/6
4/6
3/6
4/6
4/6
4/6
6/6
5/6
5/6
3/5
3/6
3/6
3/6
5/6
4/6
3/5
4/5
6/6
6/6
The column “Areal ratio” indicates the ratio between the number of
areas where the signals are independently found with respect the total
number of areas with minimum coverage. RR stands for “risk ratio.” Family
counts come from Glottolog (31).
Blasi et al. www.pnas.org/cgi/content/short/1605782113
5 of 7
Table S4. Dependencies between signals involving the same
concept
Concept
Symbol 1
Symbol 2
Effect
Families tested
Bone
Bone
Breasts
Breasts
Breasts
Breasts
Breasts
Breasts
Breasts
Breasts
Breasts
Breasts
Dog
Dog
Full
Full
I
I
I
I
Knee
Knee
Knee
Knee
Leaf
Leaf
Leaf
Leaf
Name
Name
Nose
Nose
Nose
Nose
Nose
Nose
One
One
Tongue
Tongue
Tooth
Tooth
We
We
We
You
You
You
You
You
You
You
y
k
h
u
u
a
a
a
m
m
m
r
s
t
b
p
b
s
t
u
k
o
q
u
l
b
b
p
o
i
a
a
n
n
u
u
n
t
E
e
b
m
l
n
n
d
r
u
o
o
o
t
k
y
m
a
m
u
m
r
h
u
a
a
t
s
p
b
t
u
b
s
q
u
k
o
b
l
p
b
i
o
n
u
a
u
a
n
t
n
e
E
m
b
n
l
p
t
o
o
r
s
u
d
−0.04
−0.14
−0.04
−0.16
−0.10
−0.16
0.18
0.11
−0.10
−0.08
0.12
0.03
−0.08
−0.04
−0.17
−0.21
0.02
−0.02
0.04
−0.07
−0.19
−0.28
−0.22
−0.29
0.10
0.09
−0.18
−0.21
−0.06
−0.12
0.05
−0.09
0.05
−0.05
−0.09
−0.06
−0.07
−0.06
−0.17
−0.16
0.03
0.02
−0.04
−0.19
−0.06
−0.04
0.02
−0.10
0.15
−0.04
−0.11
−0.04
298
298
237
314
309
314
317
285
237
309
317
285
281
281
175
175
264
265
264
265
71
273
71
273
217
217
226
226
290
290
329
321
329
319
321
319
338
338
142
142
272
272
257
257
279
253
254
285
254
252
285
253
The effect is the genealogically balanced mean change in probability of
finding the first symbol given that the second is present (as estimated by the
mixed model). Only entries with q values smaller than 0.05 shown. See Materials and Methods for further details.
Blasi et al. www.pnas.org/cgi/content/short/1605782113
6 of 7
Table S5. Spatial and population analysis
Distance to nearest neighbor
Population
Concept
Symbol
β
diff. ð0−1,000Þ
P value
β
diff. ð1−10,000Þ
P value
Stone
Full
Dog
Tongue
Knee
Knee
Nose
Fish
Knee
Leaf
Leaf
Name
One
Star
Tongue
Tongue
t
p
s
E
o
u
n
a
p
b
p
i
t
z
e
l
−0.591
−0.542
−0.441
−0.357
−0.263
−0.259
−0.244
—
—
—
—
—
—
—
—
—
−0.296
−0.444
−0.182
−0.343
−0.271
−0.249
−0.196
—
—
—
—
—
—
—
—
—
0.009
0.005
0.046
0.034
0.030
0.023
0.036
—
—
—
—
—
—
—
—
—
—
—
0.787
—
—
—
—
1.009
−1.087
0.574
−0.506
−0.420
−0.575
0.864
−0.358
1.126
—
—
0.057
—
—
—
—
0.176
−0.121
0.055
−0.052
−0.077
−0.063
0.054
−0.059
0.175
—
—
<10−3
—
—
—
—
<10−3
<10−3
0.007
0.042
0.008
0.002
0.049
0.028
<10−3
Estimated parameters (β), genealogical balanced mean probability difference (diff), and P values for the
distance to nearest neighbor (dnn) and population model, displayed only for the signals and variables that
reached significance at α = 0.05. See main text for details.
Table S6. Analysis of word-initial position bias
Concept
Symbol
Bias
Lineages
Bite
Bone
Breasts
Breasts
Ear
Fish
Full
Full
Horn
Horn
Knee
Knee
Knee
Knee
Leaf
One
Red
Tongue
k
k
u
m
k
a
b
p
r
k
o
p
k
q
l
n
r
l
0.20
0.09
−0.06
0.05
0.07
0.05
0.11
0.12
−0.23
0.15
0.10
0.09
0.07
0.19
−0.14
−0.07
−0.24
−0.09
42
162
185
152
159
249
81
100
82
115
177
104
177
35
120
175
28
160
Bias measure how more or less frequently the symbol appears in word
initial position for that concept. Lineages counts how many lineages had at
least one language for which the analysis could be performed. See Materials
and Methods for more details.
Blasi et al. www.pnas.org/cgi/content/short/1605782113
7 of 7